Cuestionario vinculado a una guía · Medio Nivel

Prueba sobre la herramienta de reclutamiento de IA desechada de Amazon

Prueba cómo el modelo de currículum de Amazon aprendió sesgos de las etiquetas de contratación históricas, por qué falló la eliminación de palabras de género y cómo se filtran las variables proxy.

Rutas de guía relacionadasAmazon Caso de sesgo de la herramienta de reclutamiento de IA
Pregunta 1 de 8

What data was Amazon's recruiting model trained on?

Vista previa de cada pregunta8 preguntas

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

  1. What data was Amazon's recruiting model trained on?
    • Anonymous public job postings from competitors
    • Résumés submitted to Amazon over a ten-year period
    • Synthetic résumés generated to be gender-balanced
    • Performance reviews of current senior executives only
  2. Which word did Reuters report the model penalized?
    • "Manager"
    • "Executed"
    • "Captured"
    • "Women's"
  3. Why did editing the model to ignore specific gendered terms fail to solve the problem?
    • There was no guarantee the model would not find other proxies to sort candidates in discriminatory ways
    • The edits made the model reject every applicant
    • The law forbade any changes to the model
    • The edits deleted all training data
  4. What is the central lesson about the training label in this case?
    • Labels should always be generated at random
    • The model had too few parameters to learn fairly
    • The label reflected past hiring decisions, which contained bias
    • Labels do not affect what a model learns
  5. Which verbs did the model reportedly favor because they appeared more often on men's résumés?
    • "Led" and "built"
    • "Executed" and "captured"
    • "Designed" and "coded"
    • "Managed" and "supported"
  6. What is the selective labels problem in hiring models?
    • Recruiters choose which labels to delete
    • Models can only read résumés shorter than one page
    • Only certain job titles are allowed as labels
    • Outcomes are observed only for people who were hired, so rejected candidates' potential is never measured
  7. Under the four-fifths rule of thumb, what signals possible adverse impact?
    • One group's selection rate is below 80 percent of the highest group's rate
    • More than four-fifths of applicants are rejected
    • A model uses more than five features
    • Fewer than four in five recruiters approve the tool
  8. What does New York City's Local Law 144 require of employers using automated employment decision tools?
    • A ban on all AI in hiring
    • Paying candidates who are rejected by software
    • An independent bias audit and a published summary of results
    • Registering every applicant with the city